See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220349625

A Visited Item Frequency Based Recommender System: Experimental Evaluation

and Scenario Description

Article  in  JOURNAL OF UNIVERSAL COMPUTER SCIENCE · January 2011

Source: DBLP

CITATIONS READS
0 110

5 authors, including:

Roberto Konow Javier Pereira

Universidad Diego Portales Tecnológico de Monterrey, Campus Toluca, México
8 PUBLICATIONS   29 CITATIONS    32 PUBLICATIONS   152 CITATIONS   

SEE PROFILE SEE PROFILE

Nelson Baloian
University of Chile
182 PUBLICATIONS   880 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

FONDECYT Regular nº1161200, Co-Investigador "Simultaneous development of "21st Century ICT Skills" and curriculum content through mobile devices, using
microblogging and geocollaboration" View project

Azimut View project

All content following this page was uploaded by Nelson Baloian on 02 June 2014.

The user has requested enhancement of the downloaded file.

 A visited item frequency based recommender system:
experimental evaluation and scenario description

Roberto Konow
Informatics and Telecommunications Engineering School
Universidad Diego Portales, Santiago de Chile
roberto.konow@mail.udp.cl

Wayman Tan, Luis Loyola

SkillUpJapan Corporation
Tokyo, Japan
loyola@skillupjapan.co.jp

Javier Pereira
Informatics and Telecommunications Engineering School
Universidad Diego Portales, Santiago de Chile
javier.pereira@udp.cl

Nelson Baloian
Department of Computer Science
Universidad de Chile, Santiago de Chile
nbaloian@dcc.uchile.cl

Abstract: There has been a continuous development of new clustering and prediction
techniques that help customers select products that meet their preferences and/or
needs from an overwhelming amount of available choices. Because of the possible huge
amount of available data, existing Recommender Systems showing good results might
be difficult to implement and may require a lot of computational resources to perform
in this scenario. In this paper, we present a more simple recommender system than the
traditional ones, easy to implement, and requiring a reasonable amount of resources
to perform. This system clusters users according to the frequency an item has been
visited by users belonging to the same cluster, performing a collaborative filtering
scheme. Experiments were conducted to evaluate the accuracy of this method using
the Movielens dataset. Results obtained, as measured by the F-measure value, are
comparable to other approaches found in the literature which are far more complex to
implement. Following this, we explain the application of this system to an e-content
site scenario for advertising. In this context, a filtering tool is shown which has been
developed to filter and contextualize recommended items.
Key Words: Recommender System, Collaborative Filtering, Clustering, TF-IDF, F-
Measure, Advertising, e-content
Category: H.3.1, H.1.m, H.4.m, J.0.m
1 Introduction

Nowadays, when e-commerce has penetrated almost all branches, it becomes

important to present each one of the millions of potential customers with a
personalized offer. In this scenario, Recommender Systems play an important
role. For more than a decade, there has been continuous development of new
clustering and prediction techniques that help customers select products that
meet their preferences and/or needs from an overwhelming amount of available
choices [Sarwar et al. 2002]. Examples of those applications include recommen-
dation systems for buying books, CDs and other products at Amazon, recom-
mendation of movies to be seen at Netflix and recommendations for listening to
certain types of music at Last.fm.
During the last five years we have seen a gradual market shift on recom-
mender systems from electronic commerce to content streaming and general
media delivery, including music and movies. There are many media delivery
companies that are making efforts to improve their Recommendation Systems.
One particular example is that of Netflix, the online DVD rental pioneer in the
US, which offers a 1 million dollar prize to anyone contributing to improve its
movie recommender system Cinematch. Over the last 2 years we have seen a soar
in video-on-demand and IP-based television (IPTV) services. According to the
US IPTV Forecast and Outlook report from Strategy Analytics, it is expected
that IPTV revenues will grow rapidly, reaching 14 billion US dollars in 2012
up from 694 million US dollars in 2007 [Piper 2010]. Only in 2008, the global
IPTV market grew 63% while the US market saw a bad year due to the global
economic downturn, according to the Broadband Forum [Broadband 2010], a
worldwide consortium of around 200 companies from the telecommunication and
information technology sector. In [Cotriss 2009] and [Van den Dam 2007] we also
see pertinent background information supporting the thesis that advertising in
IPTV will become an important business in the near future.
One of the main revenue sources in the IPTV industry is expected to be
advertisement and more specifically, customer targeted advertisement. However,
implementing effective advertising in a real scenario faces some technical chal-
lenges due to the huge amount of advertised items available for showing. This
means, that a small portion of them might be shown while a user is viewing a
movie or just visiting the site. In this scenario recommender systems have an im-
portant role to play selecting the right subset of advertisement items to watch a
user will most likely react. However, traditional recommender systems will most
probably face serious complications in a real scenario since the amount of data
concerning users as well as advertisement items is huge and sometimes not very
accurate. On the other hand, there are some basic data that might help to dras-
tically filter the number of suitable advertisement items for a certain user with
a very simple method, like defining a the target user for a certain advertising
item to be delimited by age, gender or geographical area.
In this paper we present a solution implemented for a real case (an IPTV
company in Japan). Here both were applied, an automatic hybrid recommenda-
tion algorithm based on clustering the user pool according to their preferences
and a collaborative filtering process according to their reactions to previous rec-
ommendations. This is followed by a semi-automatic filtering process based on a
profile definition of target user previously defined by the advertiser. The process
is shown in Figure 1 and consists of 6 steps.
The rest of the paper is organized as follows: The next section describes the
pertinent state of the art for recommender systems. The third section presents a
recommender system tailored for a general scenario with large amounts of users
and potential items to recommend. In this chapter the user clustering and the
implicit collaborative filtering process are described in detail. An evaluation of
this automatic recommender process is presented using the Movielens database.
The aim of this evaluation was testing the suitability of the process for predicting
items (in this case movies) a user belonging to a certain cluster will chose. After
this, the fourth section shows the tools developed in order to obtain strategic
information about potential users in order to help advertisers define the filters
they would like to apply to the advertising items they provide.

Obtaining Data Clustering Collaborative Filtering Recommendation Filtering Final Recommendation

Step 1 Step 2 Step 3 Step 4 Step 5 Step 6

Automatic Processing Human Assisted

Figure 1: Process diagram of the system

2 Research Background

Personalization [Candillier et al. 2008] consists of gathering, storing and an-

alyzing information [Schirru et al. 2010] about visitors of a web site or sys-
tem in order to deliver the right information to each visitor at the right time
[Arunachalam and Thambidurai 2010]. Personalization meaning the ability a sys-
tem has to recommend items a user might find interesting. In this sense, a
recommender system may offer particular types of personalization mechanisms
[Manouselis and Costopoulou 2007]:

– managing the information overload [Maes 1994, Klein et al. 2006],

 – aiding to detect the user’s preferences,

– relating them to predicted preferred items,

– filtering them from non-interesting or non-relevant responses.

From a process point of view, a recommendation is a response to a user

request where an inference task produces a list of items credibly correlated or
associated with the user preferences. Formally, let U be a set of users, I a set of
items and v(u, i) : U ×I → < a value function, or rate, measuring the explicit (or
implicit) preference of a user u ∈ U for an item i ∈ I. Hence, the user-item data
structure defined as V = [v(u, i)]u∈U,i∈I corresponds to the matrix containing
the rates of users for items.
Regularly, a recommender system computes the aggregated rate or predicted
value of an active user for a given item. Based on that value, the list of most pre-
ferred items may be recommended. More precisely, the Top-N recommendation
problem may be defined as follows [Deshpande and Karypis 2004]:

Given a user-item matrix V and a set of items I that have been rated (or
viewed) by a user, identify an ordered set of items X such that | X |≤ N
and X ∩ I = ∅.

According to the process implemented to identify X, recommender systems may

be classified as: content-based [Pazzani and Billsus 2007], presenting the user
with items similar to those preferred in the past, mainly based on items features
or descriptive tags [Memmel et al. 2009]; or collaborative filtering, where items
preferred by similar users are presented to the active user; hybrid systems are also
recognized, which combine content-based and collaborative filtering approaches
[Adomavicious and Tuzhilin 2005]. Indeed, in collaborative filtering two main
approaches are usually recognized:

– Memory-based:
In memory-based algorithms, recommendations are computed based on pre-
viously rated items. The user-based algorithm class is frequently imple-
mented, which unfolds in three main steps. In the first step, the most simi-
lar users, as compared to the active one, are identified. Regular techniques
may be used to compute similarity between pairs of rating vectors in V
[Choi et al. 2010]: Pearson correlation, Jaccard Pearson or the cosine sim-
ilarity, among others. In the second step, an active user’s neighborhood is
discerned, based on the similarity measure. Classical methods of doing this
are center-based neighborhood, K-Nearest Neighbor and clustering. In the
third step, a list of recommendations, ordered by the predicted value, is pre-
sented. The value v(u, i) may be calculated as the simple average or the
 weighted sum of ratings for items evaluated by nearest neighbors, not rated
by the active user. Although the user-based approach is very popular, it has
two documented drawbacks. First, the low performance in contexts of high
number of items/users and sparsity of matrix V , and the “cold-start” prob-
lem (when no ratings are available for a user interacting for a first time with
a recommender system. [Schein et al. 2002]).

– Model-based:
In model-based approaches, a model derived from the analysis of avail-
able data is used to predict the v(u, i) values [Sarwar et al. 2002] . This
is an “off-line” process, updating the model every time enough changes on
V have occurred. One implementations of this approach is that users are
clustered into classes such that an item rating is predicted from ratings in
a class. Several techniques have been implemented for clustering purposes
[Sandvig et al. 2008]: K-Nearest Neighbor, k-Means clustering, probabilistic
Latent Semantic Analysis or Principal Component Analysis, among others
[Adomavicious and Tuzhilin 2005]. In some cases, the item-based technique
is usually implemented, where predicted ratings are based on items correla-
tions instead of users’ similarities. It has been argued that if the item-based
method is less dynamic than the user-based method, then a model may
be constructed [Deshpande and Karypis 2004]. However, in this approach
model obsolescence should be considered since changes may affect the accu-
racy of the recommendations.

The recommender system proposed in this article , which is based on the

work presented in [Konow et al. 2010], may be classified as an hybrid one as
it has characteristics of both model types: It can be considered a model-based
approach since it does cluster users according to demographics and user’s prefer-
ences for movie categories. In addition, aggregated rates are calculated on-line,
in a memory-based method, considering users with similar preferences. Practical
reasons justify this model. First, assuming that the user’s preferences for movie
categories are relatively stable, there is no need for frequent user clustering,
a process which takes time and resources. Second, instead of maintaining the
preferences vector for each user we maintain one for the whole cluster. Clearly,
usefulness of our model assumes that clusters are correctly defined and the near-
est neighbors are detected.
When selecting a recommender system algorithm, properties affecting the
user experience need to be identified [Shani and Gunawardana 2011]. Conse-
quently, different techniques exist for evaluating recommender systems, depend-
ing on the recommendation purposes [Hernandez and Gaudioso 2008]. On one
hand, a recommender system may be evaluated by metrics according to the Infor-
mation Retrieval research area: recall, precision and ROC. Thus, recall measures
the coverage of useful items proposed by the recommender system, while pre-
cision measures the capacity of the system to show only relevant items among
recommendations presented to the active user. On the other hand, when items
are rated or a measure is assigned to them (for instance, the number of times an
item has been chosen), error is usually measured as the capacity of the system
to correctly predict the value that a user would assign to an item not yet eval-
uated. Common measures are mean absolute error (MAE), mean square error
(MSE) or root mean square error (RMSE), among others. In addition, different
techniques and metrics exist for evaluation in terms of the inferred and real rat-
ings (representing the real interests of a user): Spearman correlation, normalized
distance-based performance measure or Half-life utility [Buriano et al. 2006]. As
will be explained later, our recommender system’s quality must be evaluated
using metrics taking into account the type of rate used to evaluate items. In this
case, the most promising results have been obtained when a “most viewed” rate
is used.

3 The proposed Recommender System

Let us consider that the automatic recommendation process is triggered by a

user request and obeys to the following steps (see Figure 1):

1. Identification of the active user’s preferences (obtaining and representing

data).

2. Searching for similar users matching these preferences (clustering).

3. Inference of credibly most preferred items (collaborative filtering).

4. Presentation of preferred items and/or preference values (final recommenda-

tion).

Step 1 usually assumes a data structure to represent both a user and his/her
preferences. In the case of memory-based approaches the data is compiled into
the V matrix. Differently, in the model-based side of our system, personal data
such as genre, age and the user’s preference for movie categories are used as
the information base for clustering purposes. Knowing the cluster where the
active user belongs, Step 2 consists of a systematic identification of similar users
inside, using the K-Nearest Neighbor algorithm. In order to select the distance
measure that best performs in our model, this algorithm has been tested with
different metrics: Pearson, Jaccard Pearson, cosine and Euclidean (see Section
3.3). In Step 3, the most suitable items are searched among items rated by
similar users and the aggregated rate is computed for each one of them. Step 4
corresponds to the recommendation stage, where the Top-N items are listed. Let
us consider the scenario in where a user watching a movie may simultaneously
receive a recommendation concerning other movies, coming from the service
provider. The main hypothesis behind this work is that people with similar
preferences for a certain movie genre and similar profile characteristics may have
similar preferences (negative or positive) for movies. In fact, there are reasonable
arguments supporting this hypothesis. For example, we might expect that people
watching frequently musical movies will be attracted to follow links for a concert
or pop music movies. Moreover, if two persons watching the same movie are
also in the given age range, there is a high possibility that they like the same
music and hence would follow the same items. In the same way, people frequently
watching cooking programs on the TV may be also interested in programs about
restaurants, or programs where special cooking recipes are shown.

3.1 Clustering users

Many recommender systems find relationships between clusters of users and

clusters of items. A newcomer to the system is classified into a cluster in order
to present her/him with the most relevant items for the cluster. Given the hy-
pothesis that people watching similar movies might have common preferences
and hence follow similar links, we group the users according to the genre of the
movies they already watched combined with their age and gender.
In order to measure the preference one user has for a particular movie genre,
we adapt the Term Frequency-Inverse Document Frequency, TF-IDF, which is
commonly used in the Information Retrieval domain [Spark 1973]. Therefore, let
Mu,i be the movies of genre i watched by a user u, Mu,• the total number of
movies watched by u, M•,i the total number of movies of genre i and M the total
number of movies. Then, the preference of a user u for a genre i, Pu,i , expressed
as a TF-IDF score, may be defined as follows:

Mu,i M
Pu,i = × log( ). (1)
Mu,• M•,i
The score effectively reflects the genre preference of a user. For example, a user
who has watched several movies, each one from a different genre, will have a low
score for all genres, meaning that the user does not have a special preference
for any of them. On the other hand, a user who has watched a small number
of movies but most of them are from the same genre, will have a high score
indicating a strong preference towards that genre.
The following expression normalizes the previous value, when considering the
total sets of genres:

Pu,i
Peu,i = qP . (2)
P 2
j u,j
 Based on this measure a clustering process may be developed. The X-means
clustering algorithm [Pelleg and Moore 2000] has been implemented in this case.
In Figure 2 five clusters are shown, formed when age and preferred movie genre
are used as parameters in (2). The Y axis shows the genres of movies, the X axis
shows the age. Bullets inside a cluster indicate a similar score. These results are
used by the recommender engine to differentiate one type of user from another
based on the genre of movies they watch, thus implementing a preliminary user
behavior categorization. Utilizing a collaborative filtering approach, these results
can be used to establish an indirect relationship among users within the same
cluster.

Figure 2: Data clustering from a Movielens sample

3.2 An Implicit Collaborative Filtering method

The goal of the recommender system is selecting a limited number of movies

from a huge set, which are the most likely to be clicked by the user. The display
time during the film (timing factor), the layout for recommendations and their
number per unit of time, plays an important role in the decision of the user on
whether to click on it or not. However, we are not going to tackle those aspects
in this paper since these are issues for marketing experts, human-computer in-
terface experts and graphic designers, and can be approached independently of
(but complementary to) the recommender system itself. We hereby assume the
problem the recommender system has to solve is to choose a certain number d
of relevant movies for each user.
 When selecting the movies, the system knows that a certain user u, aged n,
of a certain gender s (male or female), who has been assigned to a cluster Cj
(1 ≤ j ≤ Nc , being Nc the number of clusters) is selecting the movie ig of a
certain genre g. We describe this situation with a pair user-movie (un,s,c , ig ). At
the beginning, the system has little information to give preference to one item
over others, so the subset of the d selected items which will be presented to
the user can be selected randomly. The user will choose some of them and this
information will be recorded.
As the system is being used, information is collected about the number of
times a user u ∈ Cj has chosen a certain recommended item Ai . This is registered
in a counter variable ciju . Thus, for each item Ai there are Na associated counter
variables, each one for a item. This information is used by the system in the
following way: a portion drec of the d alternative movies the system will present
to the user will be chosen from the ones having the highest aggregated value
P
cij = r∈Kju cijr , where Kju is the K-Nearest Neighbor of the user u. In other
words, these will be the items most frequently followed by users belonging to the
same cluster.
Another group drand of items will be chosen randomly, maintaining the equa-
tion drec + drand = d. The reason for having a certain portion of the movies
randomly chosen is to avoid the self-breeding phenomena, which will cause the
system to select always items from the set having the highest counting, which
will in turn be the only one with the possibility to increase its number of counts.
The numbers drec , drand and d are parameters of the system which may be set
according to marketing criteria (see Section 4).
The first rule to choose drec reflects the hypothesis that users from the same
cluster are likely to find interesting the same items. The second rule avoids that a
set of items are repeatedly shown because they were the first ones to be selected
by the users of the cluster. In Figure 3 a schematic view of the data structure is
shown. The elements from I1 to I6 represent the movies and U1 to U7 the users in
the active user neighborhood. We may assume that values represent the number
of times that items have been visited by users. In such a case, a 0 value means
that the user has not seen the respective item. In terms of most viewed items
by users in the neighborhood, I5 is the first item chosen by the recommender
engine to be presented to the active user. Next, I2 , I6 are to be included in the
Top-N list, etc. Depending on weather the user clicks on any recommended item,
the respective counter will be updated. Notice that a similar structure could be
used for ratings of users to items, as suggested by the schema in Figure 3.

3.3 Evaluating the recommender system model

In order to evaluate the quality [Hernandez and Gaudioso 2008] of the pro-
posed recommendation system, we conducted an experiment using the Movie-
 Cluster 1
Rating of U1 for I 2 Item I2 has not been viewed by U7
U1 U2 U3 U4 U5 U6 U7
I1 3 0 5 3 0 1 0
I2 2 4 0 2 4 2 0
I3 1 0 1 0 0 1 5
I4 0 0 0 4 1 5 3
I5 3 3 2 3 0 1 3
I6 0 2 0 1 1 5 3

Average rating for item 2 in cluster 1 = 2,8

Sum of views of item 2 in cluster 1 = 5

Figure 3: Movie evaluation schema

Lens (www.movielens.org) corpus, which is an open database of anonymous

people rating movies. It is maintained by the GroupLens community for re-
search purposes. This is an offline experiment since pre-collected data about
users choosing and/or rating data is used [Shani and Gunawardana 2011]. At
the time we conducted the test, it contained 3900 movies with 1 million rat-
ings, provided by 6040 users having rated at least 20 movies each; ratings
were integer numbers from 1 up to 5. Sparsity of this dataset was measured
as (1000000 − 1)/(6043 ∗ 3900)) = 0.04.
The experiment setting was defined as follows. First, users were clustered
according to the watched movies genre, age and gender (see Figure 2). Sec-
ond, in order to simulate items not viewed or rated by users, a testing set
was defined by randomly deleting 20% of ratings for each user. Thus, the idea
was testing if the predicted values match those items previously deleted. Third,
four K-NN neighborhood clustering process were implemented for analysis, us-
ing the following distance measures: cosine, Euclidean, Pearson, Jaccard Pear-
son. These metrics were computed by a RapidMiner software implementation
(www.rapidminer.com), which allows for an easy selection of clustering parame-
ters and algorithms. The number of neighbors, K, was set at 150 and 500. Fourth,
a Top-N strategy [Adomavicious and Tuzhilin 2005] was evaluated, defining N =
20 fixed in this setting.
The recommender system presents the user with a manageable list of relevant
movies that she/he should click in order to reach the streaming web site. Hence-
forth, in order to evaluate the recommender system the F-measure metrics was
selected, which is defined as the harmonic mean between recall and precision,
expressed as follows [Hernandez and Gaudioso 2008]:

precision × recall
Fmeasure = 2 . (3)
precision + recall
Accuracy of the recommender system was evaluated by its capacity to retrieve
relevant items among the first 20 recommended movies. Given a specific distance
measure and K, up to 33 simulations were run, each one selecting a random test-
ing set (20% of ratings). Then, the average Fmeasure value was computed. Two
independent criteria were selected for computing the Top-N items: (i) Predicted-
Rating, and (ii) Most-Viewed items. In Figure 4, the Fmeasure metric is compared
for the different distance measures, in the case of items recommended by pre-
dicted value or most viewed. Only the results for K ∈ {150, 500} are depicted.
Qualitative analysis shows that, given a distance measure, the Most-Viewed
criterion outperforms Predicted-Rating. Furthermore, the Jaccard Pearson mea-
sure is clearly the best decision in this model. Notice that, given a criteria, the
Euclidean based recommendations are outperformed in all cases.

Pearson Cosine Jaccard pearson Euclidean

0.250

0.200
F-Measure

0.150

0.100

0.050

0.000
Prediction k150 Prediction k500 Most Viewed k150 Most Viewed k500

Figure 4: F-measure for “predicted value” and “most viewed” criteria

Let F p and F v be the respective F-measures in the case of predicted ratings

(p) and most viewed (v) criteria. The hypothesis to be tested is H0 : F p = F v .
Values for results in Figure 4 are presented in Table 1. One will notice that
values for the Most-Viewed criterion are greater or equal (except for one case)
to the Predicted-Rating results, in all cases. Interestingly, the mean Fmeasure
obtained in the case of Jaccard Pearson and most viewed (0.21) is comparable
to results obtained by [Cremonesi et al. 2011] in the case of Non-Normalized
Cosine Neighborhood, Asymmetric SVD or Top Popular algorithms.
Given that four methods (distance measures) do not guarantee that statisti-
cal significance is achieved, two fictitious metrics were included. The Max-Min
row corresponds to a fictitious method having the best values on the predic-
tion criterion, but the worst ones on the viewing criterion. It is assumed that
such a method represents a conservative optimal situation in favor of Predicted-
Rating. In addition, a Mean row has been included, which corresponds to a
fictitious method reaching the mean value obtained for the real distance mea-
sures. This method takes into account a moderately acceptable situation in favor
of Most-Viewed.

Table 1: Fmeasure for different distance measures

Predicted-Rating Most-viewed
Distance
K = 150 K = 500 K = 150 K = 500
Pearson 0.001 0.001 0.040 0.090
Cosine 0.001 0.001 0.130 0.160
Jaccard Pearson 0.010 0.009 0.170 0.210
Euclidean 0.001 0.001 0.010 0.010
Max-Min 0.010 0.009 0.010 0.010
Mean 0.000 0.003 0.090 0.120

The Wilcoxon rank-sum test was applied on F-measures data contained in

Table 1. The F v − F p negative mean rank was zero both on K = 150 and
K = 500, with p = 0.043 (two-tailed) in the first case and p = 0.028 (two-tailed)
in the latter option. This means that H0 is rejected at 5% of confidence, but also
that the F v values are higher than the F p values, which indicates that the Most-
Viewed criterion prevails. That seems appropriate for our model since, first, it
is not expected that users regularly rate movies and, second, follow-up users’
activities well represent the advertising scenario presented in Section 4.

4 The advertising scenario

4.1 Introduction

In a real setting, a Japanese digital media distribution company is planning

to add advertising to its delivered video contents (movies and TV programs).
The advertising may consist of linked banners, pictures or even in-roll videos
which are displayed in multiple ways according to the web page design, including
overlays, side bars, etc. The goal of the company is to maximize the number of
times that users click on those linked ads since the ad owner pays for each visit
to its website generated from the system. The restriction is that during the
movie, only a limited number of ads can be shown from a huge available set.
This is true not only because of the limited duration of a movie, but also because
presenting too much advertising may upset the viewer or make him/her simply
ignore it [Unni and Harmond 2007]. It is then necessary to select and show those
ads which the user will most probably follow.
One of the main revenue sources in the IPTV industry is expected to be
advertisement and more specifically customer targeted advertisement. In this
scenario recommender systems have an important role to play. Recommender
systems for customer targeted advertisement are essentially an extension of the
existent algorithms for movie-based and user-based recommendation but there
are some fundamental differences: a system for customer targeted advertisement
must be able to select the best commercial considering at least four factors:

1. who is watching the movie;

2. what is the content (type of movie) being watched;

3. what is the popularity of the scene being played for both users in general
and users with a similar preference profile;

4. what is the business model being applied (e.g. maximization of revenues for
sponsors);

4.2 The Computer assisted filtering process

The more information we have about users’ background, preferences and prod-
ucts characteristics , the better a recommender system may perform [Kiewra 2005].
However, in real systems the availability of this information, as well as the ac-
curacy of it, depends on many uncontrollable factors: users might not want to
provide or upgrade their private information; detailed information about ad-
vertisement itself is difficult to obtain from the advertising providers, since the
number of advertisement items is huge. In consequence, a content-based strategy
cannot be applied due to the little information available about the content of
the items. The most suitable type of recommender system in this scenario is a
collaborative filtering one.
However, it has been shown that the computation of the recommended item
vs. user matrix takes O(U × A) [Linden et al. 2003], being U the number of
users and A the number of advertising items. This means that a filtering process
becomes necessary. Hence, let us consider that during the sign-up process users
provide personal information such as age, gender and geographical location. Be-
sides, the movie-related information stored by the system includes title, genre
(refined in sub-genres), duration, actors, director, etc. In order to embed,a first
filtering stage in the system we may ask the ad owners to include some target
metadata about the type of people the advertisement is aimed at, which matches
exactly with the information we have from the users: range of age, gender(s) and
location (city, prefecture, whole country). Additionally, the ad owner can choose
a number of film genders where the advertisement should never be included.
This may be particularly interesting when the advertiser might want to avoid
its product to be associated with a certain type of films or film content. In any
case, we expect a relatively small number of genres avoided by most ad owners
compared to the number of genres in which they would like to show their ads.
The digital company has implemented a system to acquire information di-
rectly from the users and the ad owners, but it also automatically builds up be-
havioral information for each user based on its history, for instance, by recording
what kind of movies a certain user prefers or at what time he watches them. The
system utilizes both the explicitly provided data and the generated behavioral
data to establish similarity relations among users through their unique prefer-
ences (for instance, see Figure 2).
Data has been obtained from three sources. First, the user’s profile, which the
user may or may not accurately complete during the registration process. This
data set includes gender, age, and address, among others. The second source
is the information about the movies. This data set includes, among others, the
title, cast, duration, genre and sub genre. The third source is the media access
log files which are automatically generated on the media delivery servers. The
media access log files include the time of play, stop and pause events, the user
ID, the purchased movie ID and the IP address of the user’s machine (although
this might be that of a firewall or proxy) among other parameters. In particular,
through the IP address it is possible to know the location from where each user
is connecting using a Geo-IP database service like http://www.hostip.info/ or
http://www.ipinfodb.com/. In this particular case, we have used the services of
Maxmind GeoIP because its coverage of Japanese locations is quite complete.
This information is very convenient especially when the user’s address is not
provided on the registration. The log files also include data about how long the
user watched a particular movie, including the starting point and the end point.
This data correspond in many cases to users seeking some particular scene within
the movie. This data could be very useful to display advertising before or after
the most popular scenes of a movie.
The media servers’ log files are parsed using several scripts written in Python.
Using the data mentioned in the previous section we implemented a complete
log analyzer that is used to merge the information of the access logs and the
information available in the database. The log analyzer works as the preliminary
process in order to obtain relevant information from users’ behavior.
4.3 Where to include recommendations

The display time during the film (timing factor), the layout for recommendations
and their number per unit of time plays an important role in the decision of the
user on whether to click on it or not. However, we are not going to tackle those as-
pects in this paper since these are issues for marketing experts, human-computer
interface experts and graphic designers, and can be approached independently
from (but complementary to) the recommender system itself. We hereby assume
the problem the recommender system has to solve is that of choosing a certain
number of relevant movies for each user.
Data available in log files is of paramount importance to target advertising
since it allows for the answering of questions like:

– What are the most popular movies for female/male customers between X
and Y years old living in the regions A, B and C?

– Which are the hottest scenes inside a certain movie for a given group of
users?

– Where do the majority of the young customers live?

– How many elderly customers do we have, and what do they watch?

This information certainly helps advertisers in defining the genre of the films
where the ad should not be displayed, the age of the target audience and the
geographical location where the ad should appear (see section 3.1). It also helps
the business planning division of the company to decide which type of advertising
has more possibilities of being successful and when to show it during the play
screening of a certain movie. Since the data contained in the log file is so large
it is very important to display it in an aggregated and compact way. In order
to do this, we developed a log analyzer tool, which graphically displays this
information, allowing a technical user to set the relevant parameters to filter the
information.
The main functionality of this tool displays an interface where a specialized
user has to enter the age, gender and location parameters for filtering purposes.
As an example, in Figure 5, the user has chosen to filter the data for male and
female customers between 18 and 40 years old living in Tokyo, Kyoto, Osaka,
Kanagawa, Aichi, Chiba, Saitama and Hokkaido. After this, various charts are
displayed.
 Figure 5: Filtered data from Japanese database

Following this, two pie charts showing the age distribution and gender dis-
tribution of the selected customers group are displayed (see Figure 6).

Figure 6: Distribution of age and genre of the selected group

In Figure 7 a pie chart showing the regional distribution of the customers

group is displayed, while in Figure 8 a Google Map shows the amount of cus-
tomers from each city.

Figure 7: Destribution of selected users by region

 Figure 8: Google map usage for showing user visits

Among other functionalities, the log analyzer presents statistics for a certain
movie genre or a single movie. For example, in Figure 9 we see a chart showing
the number of people who have seen a certain part of a movie. The movie is
divided into several one-minute-long pieces and the log analyzer counts how
many users have seen each piece. By looking at this report we can easily find
the scenes in the movie attracting large number of people. This information can
be used to decide when (or when not) to display advertisements.

Figure 9: Number of “Seeks” for a certain movie

5 Conclusions

In this paper we presented a recommender system which was developed to solve

the problem of recommending items to people while watching movies offered by
a IPTV Company in a real scenario. Existing recommender systems showing a
good performance require a great amount of computing power and time and are
difficult to implement because of their complexity. We developed a recommender
system based on users clustering and counting of times that users in a cluster
select advertising items. The recommendation process consists of 6 steps, which
can be grouped into two categories: full automatic and human assisted. The full
automatic set of steps comprises a users’ clustering process taking in account
their item preferences and adapting a TF-IDF score among other characteristics
like gender and age. Once the clusters are calculated, a collaborative filtering
algorithm is applied inside each cluster. It is used to select the TOP-20 rec-
ommended items for each cluster. Various K-Nearest Neighbor distance metrics
were evaluated by running several experiments using the Movielens Database.
The selected criteria for evaluating the performance of the recommender sys-
tem were two: Most-Viewed (ranking the top 20 most viewed movies from inside
the cluster) and Predicted-Rating (ranking the top 20 movies that on average
were the best rated). Results show that the most viewed criteria outperforms
using the F-Measure evaluation value. Using the Jaccard Pearson distance mea-
sure for the K-Nearest Neighbors selection and using a Most-Viewed criteria, an
average F-Measure value of 0.21 was obtained. These results can be compared
to Non-Normalized Cosine Neighborhood, Asymmetric SVD or Top Popular al-
gorithms. However, the method presented in this paper can be seen as a more
simple to understand and implement.
A log analyzer tool was implemented which is used to help company mar-
keting experts to decide whether to apply “filters” to advertisings in order to
select the population it should or should not reach, like gender, age, and/or ge-
ographical location. This tool also helps deciding in which part of the movie the
advertising item should be shown. This human assisted filtering process needs
to be evaluated and is part of our future work.
The main contribution of this work is a recommender system which performs
as good as others in the literature, but being much simpler to understand and
implement. It uses only the strict necessary information to compute results, sav-
ing computational as well as human resources. It also tackles the cold start and
over-fitting problems, very common in recommender systems. However, future
tests need to be done in order to cope with performance aspects as computa-
tional efficiency, the effects of the choice of parameters in the model and the
clustering approach. An evaluation process on the real scenario described above
is also intended, refining our results by testing pre and post filtering provided
by the log analyzer tool.

References
[Adomavicious and Tuzhilin 2005] Adomavicius, G. and Tuzhilin, A. “Toward the next
generation of recommender systems: A survey of the state-of-the-art and possible
extensions”. IEEE Transactions on Knowledge and Data Engineering, 17, 6 (2005),
734-749.
[Broadband 2010] Broadband Forum: Strong DSL and IPTV growth,
press article released on April 2010, Last visited on September 2010
http://mybroadband.co.za/news/broadband/11908-Strong-DSL-and-IPTV-
growth.html.
[Buriano et al. 2006] Buriano, L., Marchetti, M., Carmagnola,F., Cena, F., Gena, C.
and Torre, I. “The role of ontologies in context-aware recommender systems”.
MDM’06 Proceedings of the 7th International Conference on Mobile Data Man-
agement (2006), 80.
[Candillier et al. 2008] Candillier, L., Jack, K, Fressant, F. and Meyer, F. “State-of-
the-art Recommender Systems”. In Collaborative and Social Information Retrieval
and Access: Techniques for Improved User Modeling. Eds. Chevalier, M., Christine,
J., Soulé-Dupuy, C., Information Science Reference Publisher (2008), 1-22.
[Choi et al. 2010] Choi, S. S., Cha, S. H. and Tappert, C. “A Survey of Binary Simi-
larity and Distance Measures”. Journal on Systemics, Cybernetics and Informatics,
8, 1(2010), 43-48.
[Cotriss 2009] Cotriss, D.: IPTV Advertising Comes To Life ,
http://www.dailyiptv.com/news/iptv-advertising-rises-031207/, October 2010.
[Cremonesi et al. 2011] Cremonesi, P., Garzotto, F., Negro, S., Papadopoulos, A. and
Turrin, R. “Comparative evaluation of recommender system quality”. Proceedings
of the 2011 annual conference extended abstracts on Human factors in computing
systems, Vancouver, BC, Canada, May 07-12 (2011), 1927-1932.
[Deshpande and Karypis 2004] Deshpande, M. and Karypis, G. “Item-based top-N
recommendation algorithms”. ACM Transactions on Information Systems (TOIS),
22, 1(2004), 143-177.
[Hernandez and Gaudioso 2008] Hernández del Olmo, F. and Gaudioso, E. “Evalua-
tion of recommender Systems: a new approach”. Expert Systems with applications:
An International Journal, 35 (2008), 790-804.
[Kiewra 2005] Kiewra, M. “RankFeed - Recommendation as Searching without
Queries: New Hybrid Method of Recommendation”. Journal of Universal Computer
Science, 11, 2(2005), 229-249.
[Klein et al. 2006] Klein, G., Brian M and Hoffman, R. “Making Sense of Sensemaking
2: A Macrocognitive Model”. Intelligent Systems, IEEE, 21, 5 (2006), 88-92.
[Konow et al. 2010] Konow, R., Wayman, T., Loyola,L. , Pereira, J. and Baloian,N.
“Recommender System for contextual advertising in IPTV scenarios”, Computer
Supported Collaborative Work in Design CSCWD (2010), 617-622.
[Arunachalam and Thambidurai 2010] Arunachalam, K. and Thambidurai, P. “Col-
laborative Web Recommendation Systems - A Survey Approach”. Global Journal
of Computer Science and Technology, 9, 5 (2010), 30-35.
[Linden et al. 2003] Linden, G., Smith, B. and York, J. “Amazon.com recommenda-
tions: item-to-item collaborative filter”, IEEE Internet Computing, 7, 1(2003), 76-
80.
[Maes 1994] Maes, P. “Agents that reduce work and information overload”, Commu-
nications of the ACM, 37, 7 (1994), 30-40.
[Manouselis and Costopoulou 2007] Manouselis , N. and Costopoulou, C. “Analysis
and classification of multi-criteria recommender systems”. World Wide Web: Inter-
net and Web Information Systems, Special issue on Multi-channel Adaptive Infor-
mation Systems on the WWW, 10, 4 (2007), 415-441.
[Memmel et al. 2009] Memmel, M., Kockler, M. and Schirru, R. “Providing Multi
Source Tag Recommendations in a Social Resource Sharing Platform”. Journal of
Universal Computer Science, 15, 3 (2009), 678-69.
[Pazzani and Billsus 2007] Pazzani, M. and Billsus, D., “Contet-Based Recommenda-
tion Systems”, The Adaptive web, LNCS 4321 (2007), 325-341.
[Pelleg and Moore 2000] Pelleg D. and Moore, A. “X-means: Extending K-means with
efficient estimation of the number of clusters”, Proc. 17th Int. Conf. Machine Learn-
ing (ICML”00) (2000) 727-734.
 [Piper 2010] Piper, B., “Report on IPTV Forecast and Outlook: 13.7 Billion by 2012”,
Strategy and Analytics, Feb. 27th (2008).
[Sandvig et al. 2008] Sandvig, J.J., Mobasher, B. and Burke, R. “A Survey of Collabo-
rative Recommendation and the Robustness of Model-Based Algorithms”. Bulletin
of the Technical Committee on Data Engineering, 31, 2 (2008), 3-13.
[Sarwar et al. 2002] Sarwar, B.M., Karypis, G., Konstan, J. and Riedl, J. “Scalable
neighborhood formation using clustering”. Proceedings of the Fifth International
Conference on Computer and Information Technology (2002).
[Schein et al. 2002] Schein, A., Popescul, A., Ungar, L. and Pennock, D. “Methods
and Metrics for Cold-Start Recommendations”. Proceedings of the 25th Annual
International ACM SIGIR Conference on Research and Development in Information
Retrieval (SIGIR 2002). New York City, New York: ACM., 253-260.
[Shani and Gunawardana 2011] Shani, G. and Gunawardana, A. “Evaluating Recom-
mendation Systems”. In Recommender Systems Handbook (2011), 257-297.
[Spark 1973] Spark Jones, K. “A statistical interpretation of term specificity and its
application in retrieval”. Journal of Documentation 28, 1(1973), 11-21.
[Schirru et al. 2010] Schirru, R., Baumann, S., Memmel, M. and Dengel, A. “Extrac-
tion of Contextualized User Interest Profiles in Social Sharing Platforms” Journal
of Universal Computer Science, 16, 16 (2010), 2196-2213.
[Unni and Harmond 2007] Unni, R. and Harmond, R., “Perceived Effectiveness of
Push vs. Pull Mobile Location-based Advertising”. Journal of Interactive adver-
tising, 7, 2(2007), 28-40.
[Van den Dam 2007] Van den Dam, R. “IPTV ADVERTISING: a gold mine for tel-
cos?” last visited October 2010 http://broadcastengineering.com/RF/ broadcast-
ing iptv advertising gold/index.html.

View publication stats